Skip to content

DATA.md

Latest commit

 

History

History
69 lines (51 loc) · 3.02 KB

DATA.md

File metadata and controls

69 lines (51 loc) · 3.02 KB
Raw

Data

Everything needed to reproduce the figure and the reported metrics is already in this repository — results/evaluation_data.parquet holds the label and score for every observation in every evaluated (model, dataset) pair, so python reproduce.py works from a fresh clone with no downloads.

The large binaries below are only needed to score new peptides, or to recompute the metrics from the original prediction tables rather than from the committed extract.

Zenodo deposit

Archived at DOI 10.5281/zenodo.21641350.

Archive Size Contents SHA-256
checkpoints.zip 2.3 GB 30 model checkpoints (models/) 572eb0e5b8fd90044fe50c284a5e69969de12dace006731fc90a9ce35bb71828
datasets.zip 393 MB 85 dataset bundles (datasets/) a654b2ef29d27297124f69ab2b50cf720ba254b2dccf1c18509bf78845fa0b3d
predictions.zip 548 MB 209 prediction tables (predictions/) b470dc01198a7dd71abe472bd6967b6e170b3fdd0af5cbd9fcefc538e71f550b

Unpack any of them at the repository root; each archive already contains its top-level directory, so the paths land where the code expects them:

unzip checkpoints.zip     # -> models/v1_iter0_MHCRANK-9287.ckpt, ...
unzip datasets.zip        # -> datasets/validation/golden_val_v1_fold0/, ...
unzip predictions.zip     # -> predictions/golden_val_v1_fold0__MHCRANK-9287.parquet, ...

Verify before use:

shasum -a 256 -c SHA256SUMS.txt

Per-checkpoint SHA-256 values are also in models.csv.

What is committed to git

  • all code, models.csv, datasets/datasets.csv, predictions/predictions.csv
  • results/evaluation_data.parquet (58 MB) — labels and scores for all 55 evaluated cells, which is what reproduce.py reads by default
  • results/metrics_per_version.csv, results/metrics_summary.csv, results/corruption_figure.png
  • mhc_alpha_chains.fasta

datasets/datasets.csv carries one row per bundle: role, version, kind, observation and context counts, allele count, positive rate, peptide-length distribution, and — for corrupted bundles — the model whose predictions rewrote those labels.

Bundle contents

Each dataset bundle under datasets/{training,validation}/ contains:

File Contents
observation.parquet id, seq (peptide), label, context_id
context_mhc.parquet context_id, mhc_allele
context_parent_proteins.parquet source-protein links per context
mhc_default.fasta full-length alpha chains for the alleles in this bundle
parent_proteins.fasta source protein sequences
metadata.json context and allele counts, positive rate (summarised in datasets.csv)

Model input uses the 182-residue alpha chain in mhc_alpha_chains.fasta at the repository root, not the full-length mhc_default.fasta inside the bundles.

Prediction tables are named <dataset>__<model>.parquet, with columns observation_id and value, joined to labels on observation.parquet.id.